Governed Agent Bench

A public benchmark for fail-closed behavior, delegation authority, false-success rejection, mutation receipts, and rollback discipline.

Current evidence boundary: the corpus is SAMPLE, scores are COMPUTED, and receipt checks are STRUCTURE_ONLY. A reference fixture validates the evaluator; it is not a model ranking.

Eligible model submissions

Leaderboard state

Reproducible reference

Entry	Class	Score	Cases	Corpus	Score label	Receipt verification	Model-ranked
Reference conformance fixture	SAMPLE_REFERENCE_NOT_MODEL	100	10/10	SAMPLE	COMPUTED	STRUCTURE_ONLY	false

Reference axis closure

Axis	Passed	Total	Pass rate
Non Increasing Authority	2	2	1

Immutable publication identity

{

}

Dataset · Canonical source

Model submissions remain empty until an exact JSONL trace, model and license identity, evaluator revision, result, and publication receipt are reviewed and committed together.